NFS: remove nfs_inode radix tree

The radix tree is only being used to compile lists of reqs needing commit.
It is simpler to just put the reqs directly into a list.

Signed-off-by: Fred Isaman <iisaman@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

fs/nfs/inode.c

@@ -1560,7 +1560,7 @@ static void init_once(void *foo)
 	INIT_LIST_HEAD(&nfsi->open_files);
 	INIT_LIST_HEAD(&nfsi->access_cache_entry_lru);
 	INIT_LIST_HEAD(&nfsi->access_cache_inode_lru);
-	INIT_RADIX_TREE(&nfsi->nfs_page_tree, GFP_ATOMIC);
+	INIT_LIST_HEAD(&nfsi->commit_list);
 	nfsi->npages = 0;
 	nfsi->ncommit = 0;
 	atomic_set(&nfsi->silly_count, 1);

fs/nfs/internal.h

@@ -308,6 +308,8 @@ extern void nfs_pageio_reset_read_mds(struct nfs_pageio_descriptor *pgio);
 extern void nfs_readdata_release(struct nfs_read_data *rdata);
 
 /* write.c */
+extern int nfs_scan_commit_list(struct list_head *src, struct list_head *dst,
+				int max);
 extern int nfs_generic_flush(struct nfs_pageio_descriptor *desc,
 		struct list_head *head);
 extern void nfs_pageio_init_write_mds(struct nfs_pageio_descriptor *pgio,

fs/nfs/nfs4filelayout.c

@@ -682,14 +682,16 @@ filelayout_alloc_lseg(struct pnfs_layout_hdr *layoutid,
 		int size = (fl->stripe_type == STRIPE_SPARSE) ?
 			fl->dsaddr->ds_num : fl->dsaddr->stripe_count;
 
-		fl->commit_buckets = kcalloc(size, sizeof(struct list_head), gfp_flags);
+		fl->commit_buckets = kcalloc(size, sizeof(struct nfs4_fl_commit_bucket), gfp_flags);
 		if (!fl->commit_buckets) {
 			filelayout_free_lseg(&fl->generic_hdr);
 			return NULL;
 		}
 		fl->number_of_buckets = size;
-		for (i = 0; i < size; i++)
-			INIT_LIST_HEAD(&fl->commit_buckets[i]);
+		for (i = 0; i < size; i++) {
+			INIT_LIST_HEAD(&fl->commit_buckets[i].written);
+			INIT_LIST_HEAD(&fl->commit_buckets[i].committing);
+		}
 	}
 	return &fl->generic_hdr;
 }
@@ -767,11 +769,6 @@ static const struct nfs_pageio_ops filelayout_pg_write_ops = {
 	.pg_doio = pnfs_generic_pg_writepages,
 };
 
-static bool filelayout_mark_pnfs_commit(struct pnfs_layout_segment *lseg)
-{
-	return !FILELAYOUT_LSEG(lseg)->commit_through_mds;
-}
-
 static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 {
 	if (fl->stripe_type == STRIPE_SPARSE)
@@ -780,13 +777,39 @@ static u32 select_bucket_index(struct nfs4_filelayout_segment *fl, u32 j)
 		return j;
 }
 
-struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
+/* The generic layer is about to remove the req from the commit list.
+ * If this will make the bucket empty, it will need to put the lseg reference.
+ * Note inode lock is held, so we can't do the put here.
+ */
+static struct pnfs_layout_segment *
+filelayout_remove_commit_req(struct nfs_page *req)
+{
+	if (list_is_singular(&req->wb_list)) {
+		struct inode *inode = req->wb_context->dentry->d_inode;
+		struct pnfs_layout_segment *lseg;
+
+		/* From here we can find the bucket, but for the moment,
+		 * since there is only one relevant lseg...
+		 */
+		list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list) {
+			if (lseg->pls_range.iomode == IOMODE_RW)
+				return lseg;
+		}
+	}
+	return NULL;
+}
+
+static struct list_head *
+filelayout_choose_commit_list(struct nfs_page *req,
+			      struct pnfs_layout_segment *lseg)
 {
-	struct pnfs_layout_segment *lseg = req->wb_commit_lseg;
 	struct nfs4_filelayout_segment *fl = FILELAYOUT_LSEG(lseg);
 	u32 i, j;
 	struct list_head *list;
 
+	if (fl->commit_through_mds)
+		return &NFS_I(req->wb_context->dentry->d_inode)->commit_list;
+
 	/* Note that we are calling nfs4_fl_calc_j_index on each page
 	 * that ends up being committed to a data server.  An attractive
 	 * alternative is to add a field to nfs_write_data and nfs_page
@@ -796,9 +819,14 @@ struct list_head *filelayout_choose_commit_list(struct nfs_page *req)
 	j = nfs4_fl_calc_j_index(lseg,
 				 (loff_t)req->wb_index << PAGE_CACHE_SHIFT);
 	i = select_bucket_index(fl, j);
-	list = &fl->commit_buckets[i];
+	list = &fl->commit_buckets[i].written;
 	if (list_empty(list)) {
-		/* Non-empty buckets hold a reference on the lseg */
+		/* Non-empty buckets hold a reference on the lseg.  That ref
+		 * is normally transferred to the COMMIT call and released
+		 * there.  It could also be released if the last req is pulled
+		 * off due to a rewrite, in which case it will be done in
+		 * filelayout_remove_commit_req
+		 */
 		get_lseg(lseg);
 	}
 	return list;
@@ -860,18 +888,56 @@ static int filelayout_initiate_commit(struct nfs_write_data *data, int how)
 /*
  * This is only useful while we are using whole file layouts.
  */
-static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
+static struct pnfs_layout_segment *
+find_only_write_lseg_locked(struct inode *inode)
 {
-	struct pnfs_layout_segment *lseg, *rv = NULL;
+	struct pnfs_layout_segment *lseg;
 
-	spin_lock(&inode->i_lock);
 	list_for_each_entry(lseg, &NFS_I(inode)->layout->plh_segs, pls_list)
 		if (lseg->pls_range.iomode == IOMODE_RW)
-			rv = get_lseg(lseg);
+			return get_lseg(lseg);
+	return NULL;
+}
+
+static struct pnfs_layout_segment *find_only_write_lseg(struct inode *inode)
+{
+	struct pnfs_layout_segment *rv;
+
+	spin_lock(&inode->i_lock);
+	rv = find_only_write_lseg_locked(inode);
 	spin_unlock(&inode->i_lock);
 	return rv;
 }
 
+/* Move reqs from written to committing lists, returning count of number moved.
+ * Note called with i_lock held.
+ */
+static int filelayout_scan_commit_lists(struct inode *inode, int max)
+{
+	struct pnfs_layout_segment *lseg;
+	struct nfs4_filelayout_segment *fl;
+	int i, rv = 0, cnt;
+
+	lseg = find_only_write_lseg_locked(inode);
+	if (!lseg)
+		return 0;
+	fl = FILELAYOUT_LSEG(lseg);
+	if (fl->commit_through_mds)
+		goto out_put;
+	for (i = 0; i < fl->number_of_buckets; i++) {
+		if (list_empty(&fl->commit_buckets[i].written))
+			continue;
+		cnt = nfs_scan_commit_list(&fl->commit_buckets[i].written,
+					   &fl->commit_buckets[i].committing,
+					   max);
+		max -= cnt;
+		rv += cnt;
+	}
+out_put:
+	put_lseg(lseg);
+	return rv;
+}
+
 static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 {
 	struct pnfs_layout_segment *lseg;
@@ -886,7 +952,7 @@ static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 		return 0;
 	fl = FILELAYOUT_LSEG(lseg);
 	for (i = 0; i < fl->number_of_buckets; i++) {
-		if (list_empty(&fl->commit_buckets[i]))
+		if (list_empty(&fl->commit_buckets[i].committing))
 			continue;
 		data = nfs_commitdata_alloc();
 		if (!data)
@@ -900,9 +966,9 @@ static int alloc_ds_commits(struct inode *inode, struct list_head *list)
 
 out_bad:
 	for (j = i; j < fl->number_of_buckets; j++) {
-		if (list_empty(&fl->commit_buckets[i]))
+		if (list_empty(&fl->commit_buckets[i].committing))
 			continue;
-		nfs_retry_commit(&fl->commit_buckets[i], lseg);
+		nfs_retry_commit(&fl->commit_buckets[i].committing, lseg);
 		put_lseg(lseg);  /* associated with emptying bucket */
 	}
 	put_lseg(lseg);
@@ -937,7 +1003,7 @@ filelayout_commit_pagelist(struct inode *inode, struct list_head *mds_pages,
 			nfs_initiate_commit(data, NFS_CLIENT(inode),
 					    data->mds_ops, how);
 		} else {
-			nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index], data->lseg);
+			nfs_init_commit(data, &FILELAYOUT_LSEG(data->lseg)->commit_buckets[data->ds_commit_index].committing, data->lseg);
 			filelayout_initiate_commit(data, how);
 		}
 	}
@@ -967,8 +1033,9 @@ static struct pnfs_layoutdriver_type filelayout_type = {
 	.free_lseg		= filelayout_free_lseg,
 	.pg_read_ops		= &filelayout_pg_read_ops,
 	.pg_write_ops		= &filelayout_pg_write_ops,
-	.mark_pnfs_commit	= filelayout_mark_pnfs_commit,
 	.choose_commit_list	= filelayout_choose_commit_list,
+	.remove_commit_req	= filelayout_remove_commit_req,
+	.scan_commit_lists	= filelayout_scan_commit_lists,
 	.commit_pagelist	= filelayout_commit_pagelist,
 	.read_pagelist		= filelayout_read_pagelist,
 	.write_pagelist		= filelayout_write_pagelist,

fs/nfs/nfs4filelayout.h

@@ -74,6 +74,11 @@ struct nfs4_file_layout_dsaddr {
 	struct nfs4_pnfs_ds		*ds_list[1];
 };
 
+struct nfs4_fl_commit_bucket {
+	struct list_head written;
+	struct list_head committing;
+};
+
 struct nfs4_filelayout_segment {
 	struct pnfs_layout_segment generic_hdr;
 	u32 stripe_type;
@@ -84,7 +89,7 @@ struct nfs4_filelayout_segment {
 	struct nfs4_file_layout_dsaddr *dsaddr; /* Point to GETDEVINFO data */
 	unsigned int num_fh;
 	struct nfs_fh **fh_array;
-	struct list_head *commit_buckets; /* Sort commits to ds */
+	struct nfs4_fl_commit_bucket *commit_buckets; /* Sort commits to ds */
 	int number_of_buckets;
 };
 

fs/nfs/pagelist.c

@@ -396,67 +396,6 @@ void nfs_pageio_cond_complete(struct nfs_pageio_descriptor *desc, pgoff_t index)
 	}
 }
 
-#define NFS_SCAN_MAXENTRIES 16
-/**
- * nfs_scan_list - Scan a list for matching requests
- * @nfsi: NFS inode
- * @dst: Destination list
- * @idx_start: lower bound of page->index to scan
- * @npages: idx_start + npages sets the upper bound to scan.
- * @tag: tag to scan for
- *
- * Moves elements from one of the inode request lists.
- * If the number of requests is set to 0, the entire address_space
- * starting at index idx_start, is scanned.
- * The requests are *not* checked to ensure that they form a contiguous set.
- * You must be holding the inode's i_lock when calling this function
- */
-int nfs_scan_list(struct nfs_inode *nfsi,
-		struct list_head *dst, pgoff_t idx_start,
-		unsigned int npages, int tag)
-{
-	struct nfs_page *pgvec[NFS_SCAN_MAXENTRIES];
-	struct nfs_page *req;
-	pgoff_t idx_end;
-	int found, i;
-	int res;
-	struct list_head *list;
-
-	res = 0;
-	if (npages == 0)
-		idx_end = ~0;
-	else
-		idx_end = idx_start + npages - 1;
-
-	for (;;) {
-		found = radix_tree_gang_lookup_tag(&nfsi->nfs_page_tree,
-				(void **)&pgvec[0], idx_start,
-				NFS_SCAN_MAXENTRIES, tag);
-		if (found <= 0)
-			break;
-		for (i = 0; i < found; i++) {
-			req = pgvec[i];
-			if (req->wb_index > idx_end)
-				goto out;
-			idx_start = req->wb_index + 1;
-			if (nfs_lock_request_dontget(req)) {
-				kref_get(&req->wb_kref);
-				radix_tree_tag_clear(&nfsi->nfs_page_tree,
-						req->wb_index, tag);
-				list = pnfs_choose_commit_list(req, dst);
-				nfs_list_add_request(req, list);
-				res++;
-				if (res == INT_MAX)
-					goto out;
-			}
-		}
-		/* for latency reduction */
-		cond_resched_lock(&nfsi->vfs_inode.i_lock);
-	}
-out:
-	return res;
-}
-
 int __init nfs_init_nfspagecache(void)
 {
 	nfs_page_cachep = kmem_cache_create("nfs_page",